The open-source AI startup Nous Research, supported by the crypto investment firm Paradigm, unveiled a new AI coding model called NousCoder-14B. This model was trained in just four days using 48 advanced Nvidia B200 graphics cards and claims to perform equal to or better than larger dedicated systems in certain competitions.
NousCoder-14B, developed for competitive coding, achieved an accuracy of 67.87% on the standard LiveCodeBench v6 benchmark, which is more than 7% improvement over Alibaba's base model Qwen3-14B. This success comes at a time when Anthropic’s Claude Code tool is gaining significant attention and popularity on social media.
One of the standout features of NousCoder-14B is its full open-source nature; from model weights to the reinforcement learning environment, the benchmark suite, and the Atropos training framework, allowing any researcher with access to computational resources to reproduce or extend it. This approach ensures transparency and reproducibility of research.
The training process is based on reinforcement learning with verifiable rewards, where the generated code is tested against hundreds of tests, and the model is rewarded if the code is correct. This process requires complex computational infrastructure, which Nous Research has implemented using the Modal cloud service.
A major challenge in developing AI coding models is the lack of high-quality data, specifically data involving competitive programming problems with verifiable answers. NousCoder-14B used around 24,000 problems, which accounts for nearly all data in this domain, highlighting the necessity of generating synthetic data and optimizing algorithms.
With a $65 million investment, Nous Research continues to develop open AI models and platforms, claiming that this approach enables them to compete with tech giants. They believe the future belongs to systems that not only write code but can also generate new programming problems and self-learn.